A neuro-inspired computational framework for AGI: predictive coding, active inference and free energy minimisation.
- 1Computational Psychiatry and Neuropharmacology Lab, Department of Psychology, Faculty of Health and Life Sciences, University of Exeter, Exeter, United Kingdom.
- 2School of Physiology, Pharmacology and Neuroscience, University of Bristol, Bristol, United Kingdom.
Abstract
This paper offers a conceptual synthesis of foundational principles from theoretical neuroscience [i.e., predictive coding, the Free Energy Principle (FEP) and variational inference] as a biologically grounded framework for artificial general intelligence (AGI). We adopt this predictive-coding and FEP perspective as our organising framework. It is one of several current accounts of neural computation and not all available evidence is consistent with it. These approaches characterise the brain as a hierarchical inference system that continuously updates beliefs and selects actions to minimise uncertainty and surprise. In contrast to conventional AI systems, biological agents engage in active, generative inference within dynamic, uncertain environments. We argue that this inference-based architecture and not behavioural outputs alone, plausibly underlies the adaptability, generalisation and resilience of natural intelligence. We outline a neuro-inspired computational framework built on hierarchical generative models, scalable variational inference (i.e., Variational Laplace) and Active Inference. Finally, we contrast this approach with dominant deep learning paradigms and discuss its hypothetical implications for building interpretable, adaptive and autonomous machine intelligence.
Related
- A versatile set of computational tools for inference of genetic ancestry from human molecular data
- A Covalent Organic Framework-Inspired β-Ketoenamine Crosslinking Strategy for Robust, Injectable Bovine Serum Albumin Hydrogels with pH-Triggered Drug Release
- Predictive perception via simultaneous learning and inference
- Microsatellite genotyping and computational inference for genetic management of the Eurasian wild boar (Sus scrofa L., 1758): strengths, limitations, and integrative applications
- B-GRASP: A Bayesian Framework for Inferring Graph Weights from SPDE-Inspired Dynamics
BibTeX
@article{berndt2026neuroinspired,
title = {{A neuro-inspired computational framework for AGI: predictive coding, active inference and free energy minimisation.}},
author = {Lioba C S Berndt and Alexander D Shaw},
journal = {Frontiers in human neuroscience},
year = {2026},
month = oct,
doi = {10.3389/fnhum.2026.1917153},
url = {https://europepmc.org/article/MED/42922978},
note = {CC BY},
}
RIS
TY - JOUR
TI - A neuro-inspired computational framework for AGI: predictive coding, active inference and free energy minimisation.
AU - Lioba C S Berndt
AU - Alexander D Shaw
PY - 2026
DA - 2026/10/08
JO - Frontiers in human neuroscience
DO - 10.3389/fnhum.2026.1917153
UR - https://europepmc.org/article/MED/42922978
AB - This paper offers a conceptual synthesis of foundational principles from theoretical neuroscience [i.e., predictive coding, the Free Energy Principle (FEP) and variational inference] as a biologically grounded framework for artificial general intelligence (AGI). We adopt this predictive-coding and FEP perspective as our organising framework. It is one of several current accounts of neural computation and not all available evidence is consistent with it. These approaches characterise the brain as a hierarchical inference system that continuously updates beliefs and selects actions to minimise uncertainty and surprise. In contrast to conventional AI systems, biological agents engage in active, generative inference within dynamic, uncertain environments. We argue that this inference-based architecture and not behavioural outputs alone, plausibly underlies the adaptability, generalisation and resilience of natural intelligence. We outline a neuro-inspired computational framework built on hierarchical generative models, scalable variational inference (i.e., Variational Laplace) and Active Inference. Finally, we contrast this approach with dominant deep learning paradigms and discuss its hypothetical implications for building interpretable, adaptive and autonomous machine intelligence.
SN - 1662-5161
ER -